Dataset Viewer
Auto-converted to Parquet
id
stringlengths
24
32
https://openalex.org/W426843617
https://openalex.org/W426847802
https://openalex.org/W426847806
https://openalex.org/W42684885
https://openalex.org/W42685175
https://openalex.org/W4268525
https://openalex.org/W42685308
https://openalex.org/W426853375
https://openalex.org/W426855162
https://openalex.org/W42686107
https://openalex.org/W426861303
https://openalex.org/W42686394
https://openalex.org/W42686403
https://openalex.org/W426865364
https://openalex.org/W42686726
https://openalex.org/W42686867
https://openalex.org/W426876847
https://openalex.org/W42687823
https://openalex.org/W42688086
https://openalex.org/W426883407
https://openalex.org/W426893613
https://openalex.org/W42689419
https://openalex.org/W42689598
https://openalex.org/W42689603
https://openalex.org/W42690052
https://openalex.org/W42690395
https://openalex.org/W426906
https://openalex.org/W426906305
https://openalex.org/W426909965
https://openalex.org/W4269137
https://openalex.org/W42691558
https://openalex.org/W42691641
https://openalex.org/W42692062
https://openalex.org/W426924879
https://openalex.org/W42692539
https://openalex.org/W42692553
https://openalex.org/W426925808
https://openalex.org/W42692781
https://openalex.org/W42692897
https://openalex.org/W42693242
https://openalex.org/W426935384
https://openalex.org/W42693923
https://openalex.org/W42694721
https://openalex.org/W42694722
https://openalex.org/W426954788
https://openalex.org/W42695696
https://openalex.org/W426957050
https://openalex.org/W42695855
https://openalex.org/W426959440
https://openalex.org/W4269612
https://openalex.org/W426963120
https://openalex.org/W42696518
https://openalex.org/W426966132
https://openalex.org/W42696730
https://openalex.org/W42697035
https://openalex.org/W426971244
https://openalex.org/W42697596
https://openalex.org/W42697629
https://openalex.org/W42697850
https://openalex.org/W42697960
https://openalex.org/W42698411
https://openalex.org/W42698460
https://openalex.org/W4269849
https://openalex.org/W42698543
https://openalex.org/W426985885
https://openalex.org/W426990889
https://openalex.org/W42699106
https://openalex.org/W42699327
https://openalex.org/W42699859
https://openalex.org/W427000772
https://openalex.org/W427007676
https://openalex.org/W42701455
https://openalex.org/W427015234
https://openalex.org/W42701923
https://openalex.org/W42702008
https://openalex.org/W42702547
https://openalex.org/W427025708
https://openalex.org/W42702976
https://openalex.org/W42703084
https://openalex.org/W427031700
https://openalex.org/W427033575
https://openalex.org/W42703395
https://openalex.org/W427034942
https://openalex.org/W427043642
https://openalex.org/W42704471
https://openalex.org/W42704754
https://openalex.org/W42704812
https://openalex.org/W427052601
https://openalex.org/W42705626
https://openalex.org/W42705768
https://openalex.org/W42706046
https://openalex.org/W42706136
https://openalex.org/W427062519
https://openalex.org/W42706486
https://openalex.org/W42706677
https://openalex.org/W42706740
https://openalex.org/W427071270
https://openalex.org/W42707516
https://openalex.org/W4270762
https://openalex.org/W427080033
End of preview. Expand in Data Studio

abstracts-faiss

This is a faiss index, trained on abstracts-embeddings. A ready-to-go search interface for using this index is available at abstracts-index.

Building

It was trained with the train.py script found at abstracts-search with the options -N -c 65536 (normalized, train 65536 clusters), using the default preprocess technique OPQ96_384 (PCA to a 384-dimensional vector, then apply OPQ for a 96-byte code). Note that, although the Stella model was trained with Matryoshka (MRL) loss, it outputs ordinary vectors which are not expected to be truncated, so PCA was used.

Tuning

The index comes with the Pareto-optimal parameters from faiss.ParameterSpace.explore at index/params.json, so a point on the speed-recall tradeoff can be immediately picked. For reference, the exec_time field is seconds-per-query on an i7-12700H, and the recall field is 1-Recall@1. (1-Recall@1 is the probability that the top result returned by the index is the true closest, measured using a holdout set.) The best recall is 0.756 with a search time of 0.11 seconds, but a recall of 0.723 can be had with a search time of 0.0042 seconds.

Downloads last month
47

Space using colonelwatch/abstracts-faiss 1